Towards a Universal Recommender System: a Linked Open Data Approach

Abstract

Recommender systems (RSs) play a crucial role in helping users make informed decisions in the face of an ever-increasing array of choices. Most RSs are domain-specific, but domain-independent (or general purpose) RSs can be applied to a wide range of application areas, leveraging data and insights from different domains and handling large user populations. Current RS research is dominated by "improve the state-of-the-art", in terms of accuracy, speed and scale. Truly domain-independent recommender systems currently represent a gap in research, as is shown in this work.

This thesis explores the development of a general-purpose RS and the use of LOD to integrate data from different domains in an unsupervised way. Building a system that can effectively generate recommendations for any domain without prior knowledge of the data is challenging. Linked Open Data (LOD) offers a solution to this problem by enabling the integration of data from multiple domains. We explore and evaluate various unsupervised methods for acquiring and sorting through data, and using it to generate accurate recommendations. The resulting product is a truly domain-independent recommendation framework that can, in theory, be applied to a large variety of use-cases without requiring modification. Finally, this thesis suggests future research directions for building more effective general-purpose RSs.